AI data centers need different power and cooling designs because accelerators concentrate much more computing—and the electricity and heat that come with it—into individual racks. Facilities must deliver that power reliably and remove the resulting heat at the equipment, rack and building levels. The right design depends on the specific servers and site; there is no universal rack-density threshold that dictates when a facility must switch cooling methods.
How AI changes a data center’s power and heat profile
AI workloads often run on servers with multiple high-performance accelerators. Packing more of these servers into a rack raises its electrical load. Nearly all of the electricity used by IT equipment ultimately becomes heat inside the facility, so higher rack loads also increase the heat that must be carried away.
As an Amazon Associate I earn from qualifying purchases.
The International Energy Agency (IEA) described the trend in its 2025 Energy and AI report: accelerated servers are increasing data-center power density. That density matters because the facility has to deliver power and collect heat where the equipment is concentrated—not just provide enough electricity and cooling capacity for the building in aggregate.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →What a different power design has to do
A high-density AI deployment changes the demands on the path from the utility connection to the servers. It is not only a question of whether a site has enough total electricity. Power must reach the relevant racks, while supporting infrastructure helps keep service available if normal supply is interrupted.
- Plan for the IT load and its distribution. Estimate the load of the intended servers and how it is concentrated across racks. A building-wide capacity figure alone does not show whether power can be delivered where the equipment will sit.
- Include supporting equipment. Servers are only part of a data center’s electricity use. The IEA says servers average around 60% of electricity consumption in modern data centers, with the share varying by facility. Storage, networking, cooling, uninterruptible power supplies (UPS) and backup generation also contribute.
- Coordinate power and cooling plans. Changes to server density affect both electrical delivery and heat removal. The two systems need to be designed around the same expected deployment and operating conditions.
The IEA’s global outlook shows why infrastructure is part of the discussion: in its 2025 Base Case, it projected that data centers would use about 945 TWh of electricity worldwide in 2030, compared with an estimated 415 TWh in 2024. In the same scenario, the IEA attributed nearly half of the net increase from 2024 to 2030 to accelerated servers, and around one fifth to cooling and other infrastructure. These are scenario estimates, not measured shares that apply to every facility.
Why cooling must collect heat closer to dense equipment
Traditional room-level air cooling moves heat from servers into room air, then relies on facility equipment to remove it. As rack loads rise, simply moving more air through a room may no longer suit the particular server and facility design. The heat may need to be captured closer to its source and transferred to a facility heat-rejection system.
Cooling needs vary widely. The IEA’s 2025 report puts cooling at about 7% of electricity use in efficient hyperscale data centers, compared with more than 30% in less-efficient enterprise facilities. Those figures illustrate why one cooling-share assumption should not be applied to every site.
| Approach | Where heat is collected | What to assess |
|---|---|---|
| Room air cooling | Heat is transferred to the air around the servers and removed through the room’s cooling system. | Whether the server arrangement and facility can move and remove heat at the planned rack load. |
| Rear-door heat exchange | A heat exchanger at the rack’s rear captures heat from air leaving the equipment. | Rack compatibility and how the heat exchanger connects to the facility’s heat-rejection system. |
| Direct-to-chip liquid cooling | Cold plates transfer heat from supported components into circulating liquid; a coolant distribution unit (CDU) connects the equipment-side loop with facility infrastructure. | Which components are covered, rack and manifold compatibility, CDU capacity, service access and the facility-side loop. |
| Immersion cooling | Heat is collected by liquid surrounding the equipment in an immersion system. | Whether the server hardware and operating practices are designed for that system, and how heat is rejected from its loop. |
This is a description of where each approach captures heat, not a performance ranking. The available figures here do not establish an independent, apples-to-apples comparison of cost, water use, energy efficiency or lifecycle impact across these methods.
Rank #3
What liquid cooling changes—and what it does not
Direct-to-chip liquid cooling moves heat away from selected components using cold plates and a circulating coolant loop. NVIDIA describes rack-scale liquid-cooled systems and CDUs in its vendor materials. Its reference design also describes features such as redundant CDU groups and rack-level isolation; these are examples of design choices, not universal requirements for every installation.
Liquid cooling does not remove the need for facility-level heat rejection. The heat carried by the liquid still has to go somewhere, and the result depends on the full system: servers, coolant loop, CDU and site infrastructure. Climate, water availability, operating requirements and service arrangements can affect choices, but the cited material does not establish one site-independent outcome for water use or efficiency. NVIDIA’s statements about its products should be read as vendor claims rather than independent comparisons.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to choose a design for a particular site
Start with the intended compute equipment and its deployment rather than choosing a cooling technology in isolation. A practical design review should answer:
Free tools Windows power users keep installed
One-click scans. No signup required.
- What load will be installed? Identify the server configuration and expected rack-by-rack electrical demand.
- Where will heat be captured? Determine whether room air, a rear-door exchanger, cold plates or immersion is compatible with the equipment.
- How does heat leave the facility? Trace the path from the rack or cooling loop to the building’s heat-rejection equipment.
- How will equipment be maintained? Consider isolation, leak monitoring where liquid is used, redundancy, service access and replacement procedures.
- What site constraints apply? Check available power, facility layout, climate and water conditions before assuming a particular design will be efficient or practical.
Vendor reference designs can help show how a specific system is assembled, but their configurations and performance claims should not be treated as universal benchmarks. For example, NVIDIA’s reference material is relevant to its described systems; it does not establish that all AI facilities need the same CDU arrangement or liquid-cooling setup.
How large is the broader electricity demand?
Global and U.S. estimates use different geographies, dates and outlooks, so they should not be combined as if they were one forecast. The IEA estimated global data-center electricity consumption at about 415 TWh in 2024—around 1.5% of global electricity use—and projected about 945 TWh in 2030 in its 2025 Base Case. A separate IEA summary reported that data-center electricity demand grew 17% in 2025; that is a reported annual growth figure, not the same measure as the 2030 projection.
U.S. projections also vary by publication and horizon. A December 2024 U.S. Department of Energy (DOE) announcement summarizing a Lawrence Berkeley National Laboratory (LBNL) study said U.S. data-center electricity use could double or triple by 2028. A 2026 DOE resource hub, summarizing a later LBNL estimate, reported data centers could account for 11.8% of U.S. electricity use by the end of the decade, with a scenario range of 9.5% to 15.3%. These are U.S.-specific projections framed differently from the IEA’s global estimates.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Recommended Free Tools
